Lithography Metrology Techniques - Comprehensive Documentation

1. CD-SEM Metrology

Scanning Electron Microscopy Principles

CD-SEM (Critical Dimension - Scanning Electron Microscope) is the workhorse of dimensional metrology in semiconductor manufacturing. It uses a focused electron beam to image and measure feature dimensions with sub-nanometer precision.

CD-SEM Operating Conditions:
- Accelerating voltage: 300-1000 eV (low energy to minimize damage)
- Beam current: 1-50 pA (trade-off between signal and throughput)
- Landing energy: 50-500 eV at sample surface
- Scan rate: 10-100 μs per pixel
- Image resolution: 1024×1024 to 4096×4096 pixels
- Measurement precision: 0.3-0.8 nm (3σ) for modern tools

Electron Beam Interaction with Resist

Low-energy secondary electrons (SE, < 50 eV) generated near the surface provide topographic and compositional contrast. The interaction volume depends on beam energy and material:

Penetration Depth: 10-30 nm for 500-1000 eV electrons in resist, scales roughly as E^1.5
Secondary Electron Yield: δ = 0.5-2.0 electrons per incident electron, depends on material and incident angle
Signal Sources: SE1 (from beam impact), SE2 (from backscattered electrons), SE3 (from chamber/pole piece)
Contrast Mechanisms: Topography (edge enhancement), material (atomic number), voltage (surface potential)

Edge Detection Algorithms

Converting SEM images to CD measurements requires robust edge detection. Multiple algorithms exist:

Algorithm	Method	Advantages	Challenges
Threshold	Fixed % of max signal	Fast, simple	Sensitive to noise, contrast
Maximum Gradient	Peak of derivative	Physical edge location	Noise amplification
Linear Regression	Fit slopes to edge profile	Robust to noise	Model dependent
Model-Based	2D/3D profile fitting	Accurate, profile extraction	Computationally intensive

Charging Mitigation and Resist Shrinkage

Insulating resists charge under electron bombardment, distorting images. Resist also shrinks from beam-induced crosslinking:

Charging control:
- Low beam energy (< 800 eV): Reduces charge accumulation
- Charge neutralization: Flood gun injects low-energy electrons to neutralize surface
- Conductive coating: 1-2 nm Cr or Os (rare, adds process step)

Resist shrinkage:
- Shrinkage rate: 0.5-3 nm per scan for typical resists
- Mitigation: Low dose (fast scan), single-scan measurement
- Correction: Pre-measure shrinkage rate on calibration wafer

Measurement Repeatability

CD-SEM precision is characterized by repeatability (static) and reproducibility (dynamic). Key contributors to measurement uncertainty:

Shot Noise: Statistical variation in electron emission and detection, scales as 1/√N
Scan Noise: Beam position jitter, stage vibration, field distortion
Algorithm Variation: Edge detection sensitivity to signal variations
Pattern Noise: LER/LWR of the feature itself (not metrology error)
Tool Drift: Temperature, contamination, source aging over time

Automated CD-SEM: Modern fab-integrated CD-SEMs measure 50-200 sites per wafer at 200-500 wafers per day. Recipe setup, autofocus, pattern recognition, and measurement are fully automated. Typical throughput: 10-20 sites per minute including stage move and focus.

2. Scatterometry (OCD)

Optical Critical Dimension Measurement

Scatterometry (also called Optical CD or OCD) measures feature dimensions by analyzing the diffraction pattern from periodic structures. It is non-destructive, fast, and can measure buried features that CD-SEM cannot access.

Scatterometry System Specifications:
- Wavelength range: 200-900 nm (DUV to NIR) or broader
- Angles of incidence: 40-80° (variable or fixed)
- Polarizations: TE, TM, and cross-polarizations
- Measurement time: 0.5-3 seconds per site
- Precision: 0.1-0.5 nm (3σ) for CD
- Grating pitch requirement: > 400 nm (for 193nm light diffraction)

Diffraction Theory and Model Library

Scatterometry inverts the diffraction problem: given measured spectrum, infer geometry. Rigorous Coupled Wave Analysis (RCWA) computes forward model:

RCWA Forward Model:
Input: Geometry (CD, height, sidewall angle, layers), materials (n, k)
Process: Solve Maxwell equations in periodic structure
Fourier expansion + eigenvalue problem
Output: Reflectance R(λ, θ, pol) or ellipsometric angles Ψ, Δ

Inverse Problem:
minimize: ||R_measured - R_simulated(params)||²
over: params = [CD, height, SWA, ...]
Method: Levenberg-Marquardt, Newton, library search

Regression Algorithms and Parameter Correlation

The inverse problem is ill-posed: multiple geometries can produce similar spectra. Key challenges:

Correlation: CD and sidewall angle are correlated (both affect edge position), requiring multi-angle or spectroscopic data
Uniqueness: Multiple local minima may exist, library pre-computation helps find global minimum
Model Accuracy: Real structures differ from assumed trapezoid (rounding, footing, LER), introducing systematic error
Material Dispersion: Accurate n, k data essential, especially for new materials or films

Sensitivity to Profile Changes

Scatterometry sensitivity varies by parameter and wavelength:

Parameter	Sensitivity	Optimal Wavelength
CD (Top)	High	λ ~ pitch (resonance)
CD (Bottom)	Medium	Shorter λ (penetration)
Sidewall Angle	Medium	Multiple angles, pol
Height	Very High	Broad spectrum (fringes)
Film Thickness	Very High	VIS-NIR (interference)

Hybrid Metrology: Combining scatterometry with CD-SEM leverages strengths of both. Scatterometry provides high throughput and buried layer access, while CD-SEM verifies top-CD and provides independent ground truth. Multi-technique datasets improve overall accuracy.

3. Atomic Force Microscopy (AFM)

Contact vs Tapping Mode

AFM uses a nanoscale tip on a cantilever to scan surfaces with atomic-level resolution. Two primary modes exist:

Mode	Principle	Advantages	Disadvantages
Contact Mode	Tip drags on surface, deflection measured	Fast, simple	Can damage soft samples, lateral forces
Tapping (AC) Mode	Oscillating tip intermittently contacts surface	Gentle, high resolution, less damage	Slower, more complex feedback
Non-Contact Mode	Tip oscillates above surface (van der Waals)	No contact, ideal for delicate samples	Low signal, slow, difficult in air

Cantilever Dynamics and Tip Convolution

The cantilever acts as a mechanical resonator with quality factor Q = 100-500 in air, 10,000-100,000 in vacuum. Tip geometry convolves with sample topography:

Cantilever specifications:
- Spring constant k: 0.1-50 N/m
- Resonant frequency f₀: 10-500 kHz
- Quality factor Q: 100-500 (air), 10,000+ (vacuum)

Tip specifications:
- Radius of curvature: 2-20 nm (sharp tips), 20-100 nm (typical)
- Aspect ratio: 5:1 to 10:1 (height:width)
- Material: Si, Si₃N₄, carbon nanotube (CNT), diamond

Tip convolution:
Measured profile = true profile ⊗ tip shape
→ Overestimate width, miss re-entrant features

Z-Piezo Calibration and Nonlinearity

AFM uses piezoelectric ceramics for XYZ positioning. Nonlinearity, hysteresis, and creep must be calibrated and compensated:

Nonlinearity: Piezo displacement vs. voltage deviates from linear, up to 10-20% uncalibrated
Hysteresis: Position depends on voltage history, requires feedforward compensation
Creep: Time-dependent drift after step, settles with ~100 ms time constant
Calibration: Use calibrated step heights (mono-atomic steps, etched gratings) to establish Z-scale

Surface Roughness Measurement

AFM excels at quantifying surface roughness with nm-scale resolution:

Roughness metrics:
Ra (average roughness) = (1/n) Σ |zi - z_mean|
Rq (RMS roughness) = sqrt[(1/n) Σ (zi - z_mean)²]
Rz (peak-valley) = max(zi) - min(zi)

Power Spectral Density (PSD):
PSD(f) = |FFT{z(x)}|² quantifies roughness vs spatial frequency
Used to distinguish random roughness from periodic patterns

Typical resist roughness: Ra = 0.5-2 nm, Rq = 0.8-3 nm

CD-AFM: Critical Dimension AFM uses flared or boot-shaped tips to measure sidewall profiles, including re-entrant features that CD-SEM cannot resolve. Throughput is low (1-3 sites per hour) but accuracy is high, making CD-AFM the reference method for calibrating other tools.

4. Overlay Metrology

Imaging-Based Overlay Methods

Overlay metrology measures registration error between lithography layers. Imaging-based methods use bright-field or dark-field microscopy to image alignment marks:

Image-Based Overlay (IBO) Marks:
- Box-in-box: Inner box (current layer) inside outer box (previous layer)
- Frame-in-frame: Similar but with thicker frames for robustness
- Bar-in-bar (AIM): Grating-based with better precision

Measurement principle:
1. Image both layers' marks simultaneously
2. Compute centroid or edge positions via image processing
3. Overlay = offset between layer 1 and layer 2 centroids

Precision: 0.3-1.0 nm (3σ) for modern tools

Diffraction-Based Overlay (Scatterometry)

Advanced overlay metrology uses diffraction from grating targets. The diffraction signal is asymmetric when gratings are misaligned:

Diffraction-Based Imaging (DBI) Overlay:
- Gratings on both layers with designed offset (bias)
- Measure +1st and -1st order diffraction intensities
- Asymmetry: A = (I₊₁ - I₋₁) / (I₊₁ + I₋₁)
- Overlay OVL = k · A, where k is calibration factor

Advantages:
- Small target size (10×10 μm possible)
- Robust to process variations
- Precision: 0.1-0.3 nm (3σ)
Challenges:
- Requires process-specific calibration
- Sensitive to target asymmetry from etch, CMP

Overlay Mark Design and Robustness

Mark design impacts measurement quality, especially on process-heavy wafers:

Mark Size: Larger marks (80×80 μm) more robust, smaller marks (20×20 μm) save real estate
Segmentation: Multiple segments average out local process variations
Wavelength Selection: Multiple λ penetrate different film stacks, vote on overlay
Target-In-Target (TiT): Nested targets for multi-layer registration verification

Measurement Accuracy vs Precision

Overlay metrology must distinguish precision (repeatability) from accuracy (systematic error):

Error Source	Type	Magnitude	Mitigation
Tool Repeatability	Random	0.1-0.5 nm	Averaging, low noise optics
TIS (Tool Induced Shift)	Systematic	0.5-2 nm	Calibration, matching
Process Variation	Systematic	1-5 nm	Robust targets, multi-λ
Mark Damage	Systematic	2-10 nm	Process control, inspection

Overlay Budget: Total overlay = scanner contribution + process contribution + metrology contribution. For 7nm node, total budget ~2 nm (mean + 3σ) requires scanner < 1 nm, process < 1.5 nm, metrology < 0.5 nm.

5. Focus Monitors

Phase Grating Focus Monitors

Focus metrology targets are designed to have CD that varies monotonically with defocus. Phase gratings with specific pitch and orientation create this behavior:

Phase Grating Focus Monitor Design:
- Pitch: Selected such that ±1st orders are near cutoff (NA edge)
- Typical: 180-300 nm pitch for ArF scanners
- Orientation: Horizontal and vertical gratings for X and Y focus
- Measurement: CD vs. focus slope ~0.5-2 nm CD per 100 nm defocus

Phase grating advantage:
0th and ±1st orders have 180° phase difference,
creating through-pitch intensity null at best focus

Bossung Focus Analysis

Bossung curves plot CD vs. focus at multiple doses, revealing best focus, DOF, and exposure latitude:

Best Focus: Where CD curves converge (minimum sensitivity to dose)
Depth of Focus (DOF): Defocus range meeting CD spec at nominal dose
Exposure Latitude: Dose range meeting CD spec at best focus
Process Window: Area in dose-focus space meeting all specs

Bossung Curve Equation (parabolic fit):
CD(focus, dose) = CD₀ + A·(focus - best_focus)² + B·(dose - dose₀)

Best focus extraction:
Fit CD vs. focus at multiple doses
Best focus = vertex of parabola (dCD/dfocus = 0)

Typical measurement:
Focus range: ±300 nm, step 50 nm (13 points)
Dose range: ±20%, step 5% (9 points)
Total: 13 × 9 = 117 conditions (fields)

Aberration-Based Focus Metrology

Scanner aberrations (coma, astigmatism) cause field-dependent best focus shifts. Aberration-sensitive targets identify these effects:

Phase-Shifted Pupil Filters: Targets printed with different pupil regions (central, annular, quadrants)
Differential Focus: Compare best focus from different pupil-filtered patterns
Aberration Signature: Field map of focus variations reveals Zernike components
Correction: Adjust lens elements to minimize aberration-induced focus variation

Continuous Focus Monitoring: Production wafers include focus monitor targets that are measured post-litho. Statistical process control tracks focus stability over time, triggering re-qualification if drift exceeds limits (typically ±10 nm from target).

6. Dose Monitors

Resist-Based Dose Monitors

Dose monitors are features with CD highly sensitive to exposure dose. They enable real-time dose control and correction:

Dose Monitor Design:
- Feature type: Dense lines near resolution limit (high γ)
- CD vs. dose slope: 0.5-2 nm CD per 1% dose change
- Measurement: CD-SEM or scatterometry after develop

Dose extraction:
Dose_actual = Dose_nominal · (CD_target / CD_measured)^(1/γ)
where γ = resist contrast (5-15 for CAR)

Example: CD_target = 40 nm, CD_measured = 41 nm, γ = 10
Dose_actual = Dose_nominal · (40/41)^0.1 = 0.9976 · Dose_nominal
→ 0.24% dose error

Sensitivity Curves and Linearity

Dose monitor calibration establishes CD vs. dose relationship over the process window:

Calibration Wafer: Expose dose series (e.g., ±20% in 2% steps), measure CD
Sensitivity Curve: Plot CD vs. dose, typically linear near nominal dose
Linearity Range: ±10-15% dose where linear fit is valid (R² > 0.99)
Resist Contrast: γ = -1 / slope of log(dose) vs. log(CD) (from Mack model)

Process Control Sampling

Dose monitors are placed on production wafers for continuous feedback:

Sampling Strategy	Frequency	Sites per Wafer	Purpose
Every Wafer	100%	5-9 sites	Real-time dose control, trend monitoring
Hourly	1-2 wafers/hour	21-49 sites	Spatial dose uniformity verification
Daily	1 wafer/day	49-81 sites	Full dose map, trend analysis
Weekly	1 wafer/week	169+ sites	Comprehensive scanner qualification

APC (Advanced Process Control): Dose monitor measurements feed back to scanner dose control. If measured CD deviates from target, scanner adjusts dose for subsequent wafers. Typical APC loop response: 1-5 wafer delay, correction magnitude 0.5-2% dose per 1 nm CD error.

7. Defect Inspection

Bright-Field vs Dark-Field Inspection

Optical defect inspection detects particles, pattern defects, and film anomalies on wafers. Two complementary modes exist:

Mode	Illumination	Detection	Sensitivity	Applications
Bright-Field (BF)	Normal incidence	Reflected/transmitted light	High for particles > λ/4	Unpatterned wafers, particles
Dark-Field (DF)	Oblique angle (60-80°)	Scattered light only	High for small defects, edges	Patterned wafers, pattern defects

Spatial Filtering and Sensitivity

Dark-field inspection achieves high sensitivity by blocking specular reflection and collecting only scattered light from defects:

Dark-Field Optics:
- Illumination: Narrow laser beam (266-488 nm), oblique incidence
- Collection: High-NA objective (NA 0.7-0.95)
- Spatial filter: Blocks 0th order (specular), passes scattered light
- Detectors: PMT or CCD array, multiple channels for classification

Sensitivity:
- Particle detection: 20-50 nm polystyrene latex (PSL) equivalent
- Pattern defect: 30-80 nm (depends on defect type, die design)
- Capture rate: > 90% for defects > 2× sensitivity limit

Defect Classification

Detected events must be classified to separate real defects from nuisance signals:

Particle: Foreign material (polymer, metal, dust) on wafer surface
Pattern Defect: Missing, bridging, or misshapen features in resist/etch
Film Anomaly: Scratch, void, inclusion, delamination in underlying films
Edge Defect: Wafer edge contamination or chipping
Nuisance: False positives from optical artifacts, roughness, grain structure

Nuisance Filtering and ADC

Automated Defect Classification (ADC) uses machine learning to classify defects based on optical signatures:

ADC Feature Extraction:
- Size (equivalent diameter)
- Brightness (scattered intensity)
- Texture (spatial frequency content)
- Polarization response (TE vs TM)
- Multi-channel ratios (wavelength, angle, pol)

Classification algorithms:
- Rule-based: Decision trees on features
- Machine learning: SVM, random forest, neural networks
- Accuracy: 80-95% (verified by SEM review of subset)

Throughput: 60-120 wafers per hour (300 mm wafer, full surface)

Defect Review SEM: A subset of defects (typically 50-200 per wafer) are automatically re-imaged with high-resolution SEM for verification and root cause analysis. Review SEM provides defect images, precise coordinates, and composition (via EDS if equipped).

8. Film Thickness Metrology

Ellipsometry Principles

Spectroscopic ellipsometry measures film thickness and optical constants (n, k) by analyzing polarization change upon reflection:

Ellipsometry Measurement:
Incident: Linearly polarized light (p and s components)
Reflected: Elliptically polarized (phase and amplitude change)

Measured quantities:
tan(Ψ) = |rp/rs| (amplitude ratio)
Δ = δp - δs (phase difference)

For thin film on substrate:
rp, rs = f(n_film, k_film, t_film, n_substrate, k_substrate, λ, θ)
→ Invert to extract film properties

Reflectometry and Interferometry

Simpler than ellipsometry, reflectometry measures reflectance vs. wavelength. Interference fringes encode film thickness:

Interference Condition: Constructive interference when 2·n·t = m·λ, destructive when 2·n·t = (m + 0.5)·λ
Fringe Spacing: Δλ between adjacent maxima → thickness t = λ₁·λ₂ / (2·n·Δλ)
Thin Films: < 50 nm, few fringes, requires multi-angle or ellipsometry for accuracy
Thick Films: > 1 μm, many fringes, high precision possible

n and k Determination

Optical constants n (refractive index) and k (extinction coefficient) vary with wavelength. Dispersion models describe this variation:

Common dispersion models:
Cauchy (transparent materials): n(λ) = A + B/λ² + C/λ⁴
Sellmeier (dielectrics): n²(λ) = 1 + Σ[Bi·λ²/(λ² - Ci²)]
Tauc-Lorentz (semiconductors): Includes absorption edge, bandgap

Fitting procedure:
1. Measure Ψ, Δ vs. λ (or R vs. λ for reflectometry)
2. Assume model (Cauchy, Sellmeier, etc.) with parameters
3. Fit model to minimize ||Ψ_meas - Ψ_calc(params)||²
4. Extract thickness and dispersion coefficients

Multi-Layer Film Stack Fitting

Modern semiconductor devices have 10-20+ film layers. Multi-layer ellipsometry/reflectometry fits all layers simultaneously:

Layer Stack Definition: Specify sequence of materials from top to bottom
Fixed vs Floating: Some layers (e.g., Si substrate) have known n, k; others are fitted
Correlation: Similar n, k values cause parameter correlation, need constraints or prior knowledge
Uniqueness: Multi-layer fits can have multiple solutions, require good initial guess or library

Production Ellipsometry: Fab-integrated tools measure 50-100 sites per wafer in 30-60 seconds. Recipe setup is automated using process-of-record film stacks. Typical precision: 0.1-0.5 nm (3σ) for single-layer films, 0.5-2 nm for multi-layer stacks.

9. Line Edge Roughness Measurement

SEM-Based LER Metrology

Line Edge Roughness (LER) and Line Width Roughness (LWR) quantify local CD variations along a feature. CD-SEM is the primary measurement tool:

LER/LWR Measurement Procedure:
1. Image line at high magnification (100-500k×)
2. Detect left and right edges at many positions (100-1000 points)
3. Compute edge position deviations: LER_left, LER_right
4. Compute width at each position: W(x)
5. Calculate statistics:
LER = σ(edge_position) (3σ typically reported)
LWR = σ(width) = σ(W) (3σ typically reported)

Relationship: LWR² = LER_left² + LER_right² - 2·correlation·LER_left·LER_right
If edges uncorrelated: LWR = √2 · LER

Unbiased LWR and Sampling

LER measurements are biased by measurement noise and sampling. Unbiasing corrects for this:

Measurement noise contribution:
LER_measured² = LER_true² + σ_noise²

Unbiased LER:
LER_true = sqrt(LER_measured² - σ_noise²)

Noise estimation:
- Repeat measurements on same feature
- σ_noise = standard deviation of repeat measurements
- Typical: σ_noise = 0.2-0.5 nm for modern CD-SEM

Sampling requirements:
- Line length: > 500 nm (to capture low-frequency roughness)
- Sampling interval: 1-5 nm (Nyquist: < correlation length/2)
- Number of measurements: > 100 to reduce statistical uncertainty

Power Spectral Density Analysis

PSD decomposes LER into spatial frequency components, revealing roughness sources:

PSD Calculation:
PSD(f) = |FFT{edge(x) - mean(edge)}|² / L
where L = line length, f = spatial frequency

PSD Characteristics:
- Low frequency (< 1/100nm): Process variations, scanner aberrations
- Mid frequency (1/100nm to 1/10nm): Resist chemistry, acid diffusion
- High frequency (> 1/10nm): Photon shot noise, molecular roughness

PSD integral:
LER² = ∫ PSD(f) df (Parseval's theorem)
Allows attribution of LER to frequency bands

Sampling Statistics and Confidence

LER varies from line to line. Adequate sampling is necessary for process monitoring:

Within-Die Variation: Measure 5-10 lines per site, report mean and std dev
Across-Wafer Variation: Measure 9-21 sites per wafer, map spatial trends
Wafer-to-Wafer Variation: Measure 3-5 wafers per lot, track process stability
Confidence Interval: 95% CI = mean ± 1.96·(σ/√N), where N = number of measurements

LER Specification: Advanced nodes (7nm and below) specify LER < 2 nm (3σ) for critical layers. Achieving this requires optimized resists, controlled acid diffusion, low dose noise, and etch processes that don't degrade resist LER. LER reduction is a major R&D focus for sub-5nm nodes.

10. Metrology Matching

Tool-to-Tool Correlation

Multiple metrology tools of the same type must give consistent measurements (matching). Differences cause production issues when wafers move between tools:

Matching Procedure:
1. Select golden wafer with stable, representative features
2. Measure on Tool A (reference) at many sites (25-49)
3. Measure on Tool B at same sites
4. Compute differences: Δ(site) = CD_B - CD_A
5. Calculate statistics:
Mean offset = mean(Δ)
Range = max(Δ) - min(Δ)
Std dev = σ(Δ)

Matching criterion:
|Mean offset| < 0.5 nm
Range < 1.5 nm
Std dev < 0.5 nm

Golden Wafer Reference Standards

Golden wafers serve as transfer standards between tools and over time. Requirements:

Stability: Features must not change over months (store in N₂, avoid light, temperature extremes)
Representativeness: Feature types and sizes match production (lines, contacts, various pitches)
Site Coverage: 25-81 sites across wafer to capture field and radial variations
Certification: Initial measurement on calibrated reference tool (e.g., CD-AFM)
Recertification: Every 3-6 months to verify stability, update if drift detected

Total Measurement Uncertainty (TMU)

TMU combines precision (repeatability), matching (reproducibility), and accuracy (systematic error):

TMU components:
1. Precision: σ_precision from repeat measurements on same site
2. Matching: σ_matching from tool-to-tool comparison
3. Accuracy: offset from reference method (e.g., CD-AFM)

Combined TMU:
TMU² = σ_precision² + σ_matching² + accuracy²

Gauge R&R requirement:
TMU < 10% of process tolerance

Example (CD tolerance ±3 nm):
Required TMU < 0.3 nm
→ σ_precision < 0.2 nm, σ_matching < 0.2 nm, accuracy < 0.15 nm

GR&R Studies and Variance Decomposition

Gauge Repeatability and Reproducibility (GR&R) studies partition total variance into sources:

Variance Source	Symbol	Definition
Part (Wafer) Variance	σ²_part	True wafer-to-wafer variation
Repeatability	σ²_repeat	Within-tool precision (same operator, same part)
Reproducibility	σ²_reprod	Tool-to-tool or operator-to-operator variation
Total Variance	σ²_total	σ²_part + σ²_repeat + σ²_reprod

GR&R acceptance: σ²_measurement / σ²_total < 10% (measurement variance < 10% of total variance)

Metrology Fleet Management: Large fabs have 5-20 metrology tools of each type. Monthly matching ensures fleet uniformity. Tools that drift out of spec are recalibrated or removed from production until repaired. Centralized database tracks all measurements for process control and yield analysis.